using composite of dummies instead of dummies as predictor: huge difference?

38 views
Skip to first unread message

M.

unread,
Oct 29, 2014, 9:53:43 AM10/29/14
to lav...@googlegroups.com
Dear all,

I have data collected at 18 different plots. I have created 17 dummy (0/1) variables to reflect this (F2,F3,...,F18).
In the model below, the y variables are continuous and the x variables are also dummy variables, representing another categorical exogenous variable.

model1 <- '
    # regressions
        y1 + y2 + y3 ~ x1 + x2 + x3 + x4 + F2 + F3 + F4 + F5 + F6 + F7 + F8 + F9 + F10 + F11 + F12 + F13 + F14 + F15 + F16 + F17 + F18
        y4 ~ y1 + y2 + y3 + x1 + x2 + x3 + x4 + F2 + F3 + F4 + F5 + F6 + F7 + F8 + F9 + F10 + F11 + F12 + F13 + F14 + F15 + F16 + F17 + F18
        '
This model has a good fit (ML: 2.2, df: 3,P-value: 0.53)

I then tried this:

model2 <- '
    # composite
        plot <~ 1*F2 + F3 + F4 + F5 + F6 + F7 + F8 + F9 + F10 + F11 + F12 + F13 + F14 + F15 + F16 + F17 + F18
    # regressions
        y1 + y2 + y3 ~ x1 + x2 + x3 + x4 + plot
        y4 ~ y1 + y2 + y3 + x1 + x2 + x3 + x4 + plot
        '
Now this model has a very poor fit (ML: 306.3, df: 51, P-value: 0.000), which surprised me a little.

I noticed that in model1, certain F variables have strong effects on some y variables but not at all on others. Could this be the reason for this discrepancy?
Furthermore, which model is generally preferable?

Thank you,

M.

Edward Rigdon

unread,
Oct 29, 2014, 10:46:44 AM10/29/14
to lav...@googlegroups.com

M—

      I think you are correct.  The first approach allows the weight for each dummy to vary across the outcome y variables, while the second imposes a proportionality constraint—the weights for each F must conform to the same proportionality relationship across at least y1 – y3.  In addition, the second syntax effectively defines plot as both (a) a composite sum of dummies and (b) a common factor with y1, y2 and y3 as indicators.  If there is no reason to expect y1, y2 and y3 to serve as three indicators of the same common factor, then there is no good reason to favor the second approach, I think.

--Ed Rigdon

--
You received this message because you are subscribed to the Google Groups "lavaan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lavaan+un...@googlegroups.com.
To post to this group, send email to lav...@googlegroups.com.
Visit this group at http://groups.google.com/group/lavaan.
For more options, visit https://groups.google.com/d/optout.

Martijn Vandegehuchte

unread,
Oct 29, 2014, 12:08:23 PM10/29/14
to lav...@googlegroups.com
Edward,

What you say under a) is what I was thinking.
The point you raise under b) is something I have been wondering about as well.

Is there actually any difference between specifying "y1 ~plot" or "plot =~ y1", when plot is composite or latent? You seem to suggest there isn't, right?

In textbooks I often see composites being used to sum the effects of levels of categorical variables (dummies). But then often the regressions are constructed between the composite and other latents (even with a single indicator), so no arrows directly from the composite to observed variables. Your comment b) could be the reason why.

In my model it would be difficult to only have regressions between latent variables, because of the fact that the x variables are also dummies, and I'm not sure if it makes sense to have a latent that has a single dummy variable as indicator. If not, then my model would have direct effects of these dummy x variables on the latents (again effectively making the latents into composites)...

M.





--
You received this message because you are subscribed to a topic in the Google Groups "lavaan" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lavaan/Bn7tczUsMwE/unsubscribe.
To unsubscribe from this group and all its topics, send an email to lavaan+un...@googlegroups.com.
Reply all
Reply to author
Forward
0 new messages